mm/sparse.c

   1 /*
   2  * sparse memory mappings.
   3  */
   4 #include <linux/mm.h>
   5 #include <linux/mmzone.h>
   6 #include <linux/bootmem.h>
   7 #include <linux/highmem.h>
   8 #include <linux/module.h>
   9 #include <linux/spinlock.h>
  10 #include <linux/vmalloc.h>
  11 #include "internal.h"
  12 #include <asm/dma.h>
  13 #include <asm/pgalloc.h>
  14 #include <asm/pgtable.h>
  15 #include "internal.h"
  16
  17 /*
  18  * Permanent SPARSEMEM data:
  19  *
  20  * 1) mem_section       - memory sections, mem_map's for valid memory
  21  */
  22 #ifdef CONFIG_SPARSEMEM_EXTREME
  23 struct mem_section *mem_section[NR_SECTION_ROOTS]
  24         ____cacheline_internodealigned_in_smp;
  25 #else
  26 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
  27         ____cacheline_internodealigned_in_smp;
  28 #endif
  29 EXPORT_SYMBOL(mem_section);
  30
  31 #ifdef NODE_NOT_IN_PAGE_FLAGS
  32 /*
  33  * If we did not store the node number in the page then we have to
  34  * do a lookup in the section_to_node_table in order to find which
  35  * node the page belongs to.
  36  */
  37 #if MAX_NUMNODES <= 256
  38 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  39 #else
  40 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  41 #endif
  42
  43 int page_to_nid(struct page *page)
  44 {
  45         return section_to_node_table[page_to_section(page)];
  46 }
  47 EXPORT_SYMBOL(page_to_nid);
  48
  49 static void set_section_nid(unsigned long section_nr, int nid)
  50 {
  51         section_to_node_table[section_nr] = nid;
  52 }
  53 #else /* !NODE_NOT_IN_PAGE_FLAGS */
  54 static inline void set_section_nid(unsigned long section_nr, int nid)
  55 {
  56 }
  57 #endif
  58
  59 #ifdef CONFIG_SPARSEMEM_EXTREME
  60 static struct mem_section noinline __init_refok *sparse_index_alloc(int nid)
  61 {
  62         struct mem_section *section = NULL;
  63         unsigned long array_size = SECTIONS_PER_ROOT *
  64                                    sizeof(struct mem_section);
  65
  66         if (slab_is_available())
  67                 section = kmalloc_node(array_size, GFP_KERNEL, nid);
  68         else
  69                 section = alloc_bootmem_node(NODE_DATA(nid), array_size);
  70
  71         if (section)
  72                 memset(section, 0, array_size);
  73
  74         return section;
  75 }
  76
  77 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
  78 {
  79         static DEFINE_SPINLOCK(index_init_lock);
  80         unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  81         struct mem_section *section;
  82         int ret = 0;
  83
  84         if (mem_section[root])
  85                 return -EEXIST;
  86
  87         section = sparse_index_alloc(nid);
  88         if (!section)
  89                 return -ENOMEM;
  90         /*
  91          * This lock keeps two different sections from
  92          * reallocating for the same index
  93          */
  94         spin_lock(&index_init_lock);
  95
  96         if (mem_section[root]) {
  97                 ret = -EEXIST;
  98                 goto out;
  99         }
 100
 101         mem_section[root] = section;
 102 out:
 103         spin_unlock(&index_init_lock);
 104         return ret;
 105 }
 106 #else /* !SPARSEMEM_EXTREME */
 107 static inline int sparse_index_init(unsigned long section_nr, int nid)
 108 {
 109         return 0;
 110 }
 111 #endif
 112
 113 /*
 114  * Although written for the SPARSEMEM_EXTREME case, this happens
 115  * to also work for the flat array case because
 116  * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
 117  */
 118 int __section_nr(struct mem_section* ms)
 119 {
 120         unsigned long root_nr;
 121         struct mem_section* root;
 122
 123         for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
 124                 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
 125                 if (!root)
 126                         continue;
 127
 128                 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
 129                      break;
 130         }
 131
 132         return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
 133 }
 134
 135 /*
 136  * During early boot, before section_mem_map is used for an actual
 137  * mem_map, we use section_mem_map to store the section's NUMA
 138  * node.  This keeps us from having to use another data structure.  The
 139  * node information is cleared just before we store the real mem_map.
 140  */
 141 static inline unsigned long sparse_encode_early_nid(int nid)
 142 {
 143         return (nid << SECTION_NID_SHIFT);
 144 }
 145
 146 static inline int sparse_early_nid(struct mem_section *section)
 147 {
 148         return (section->section_mem_map >> SECTION_NID_SHIFT);
 149 }
 150
 151 /* Validate the physical addressing limitations of the model */
 152 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
 153                                                 unsigned long *end_pfn)
 154 {
 155         unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
 156
 157         /*
 158          * Sanity checks - do not allow an architecture to pass
 159          * in larger pfns than the maximum scope of sparsemem:
 160          */
 161         if (*start_pfn > max_sparsemem_pfn) {
 162                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 163                         "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 164                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 165                 WARN_ON_ONCE(1);
 166                 *start_pfn = max_sparsemem_pfn;
 167                 *end_pfn = max_sparsemem_pfn;
 168         }
 169
 170         if (*end_pfn > max_sparsemem_pfn) {
 171                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 172                         "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 173                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 174                 WARN_ON_ONCE(1);
 175                 *end_pfn = max_sparsemem_pfn;
 176         }
 177 }
 178
 179 /* Record a memory area against a node. */
 180 void __init memory_present(int nid, unsigned long start, unsigned long end)
 181 {
 182         unsigned long pfn;
 183
 184         start &= PAGE_SECTION_MASK;
 185         mminit_validate_memmodel_limits(&start, &end);
 186         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 187                 unsigned long section = pfn_to_section_nr(pfn);
 188                 struct mem_section *ms;
 189
 190                 sparse_index_init(section, nid);
 191                 set_section_nid(section, nid);
 192
 193                 ms = __nr_to_section(section);
 194                 if (!ms->section_mem_map)
 195                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 196                                                         SECTION_MARKED_PRESENT;
 197         }
 198 }
 199
 200 /*
 201  * Only used by the i386 NUMA architecures, but relatively
 202  * generic code.
 203  */
 204 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
 205                                                      unsigned long end_pfn)
 206 {
 207         unsigned long pfn;
 208         unsigned long nr_pages = 0;
 209
 210         mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
 211         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 212                 if (nid != early_pfn_to_nid(pfn))
 213                         continue;
 214
 215                 if (pfn_present(pfn))
 216                         nr_pages += PAGES_PER_SECTION;
 217         }
 218
 219         return nr_pages * sizeof(struct page);
 220 }
 221
 222 /*
 223  * Subtle, we encode the real pfn into the mem_map such that
 224  * the identity pfn - section_mem_map will return the actual
 225  * physical page frame number.
 226  */
 227 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 228 {
 229         return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 230 }
 231
 232 /*
 233  * Decode mem_map from the coded memmap
 234  */
 235 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 236 {
 237         /* mask off the extra low bits of information */
 238         coded_mem_map &= SECTION_MAP_MASK;
 239         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 240 }
 241
 242 static int __meminit sparse_init_one_section(struct mem_section *ms,
 243                 unsigned long pnum, struct page *mem_map,
 244                 unsigned long *pageblock_bitmap)
 245 {
 246         if (!present_section(ms))
 247                 return -EINVAL;
 248
 249         ms->section_mem_map &= ~SECTION_MAP_MASK;
 250         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
 251                                                         SECTION_HAS_MEM_MAP;
 252         ms->pageblock_flags = pageblock_bitmap;
 253
 254         return 1;
 255 }
 256
 257 unsigned long usemap_size(void)
 258 {
 259         unsigned long size_bytes;
 260         size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
 261         size_bytes = roundup(size_bytes, sizeof(unsigned long));
 262         return size_bytes;
 263 }
 264
 265 #ifdef CONFIG_MEMORY_HOTPLUG
 266 static unsigned long *__kmalloc_section_usemap(void)
 267 {
 268         return kmalloc(usemap_size(), GFP_KERNEL);
 269 }
 270 #endif /* CONFIG_MEMORY_HOTPLUG */
 271
 272 #ifdef CONFIG_MEMORY_HOTREMOVE
 273 static unsigned long * __init
 274 sparse_early_usemap_alloc_pgdat_section(struct pglist_data *pgdat)
 275 {
 276         unsigned long section_nr;
 277
 278         /*
 279          * A page may contain usemaps for other sections preventing the
 280          * page being freed and making a section unremovable while
 281          * other sections referencing the usemap retmain active. Similarly,
 282          * a pgdat can prevent a section being removed. If section A
 283          * contains a pgdat and section B contains the usemap, both
 284          * sections become inter-dependent. This allocates usemaps
 285          * from the same section as the pgdat where possible to avoid
 286          * this problem.
 287          */
 288         section_nr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 289         return alloc_bootmem_section(usemap_size(), section_nr);
 290 }
 291
 292 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 293 {
 294         unsigned long usemap_snr, pgdat_snr;
 295         static unsigned long old_usemap_snr = NR_MEM_SECTIONS;
 296         static unsigned long old_pgdat_snr = NR_MEM_SECTIONS;
 297         struct pglist_data *pgdat = NODE_DATA(nid);
 298         int usemap_nid;
 299
 300         usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
 301         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 302         if (usemap_snr == pgdat_snr)
 303                 return;
 304
 305         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 306                 /* skip redundant message */
 307                 return;
 308
 309         old_usemap_snr = usemap_snr;
 310         old_pgdat_snr = pgdat_snr;
 311
 312         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 313         if (usemap_nid != nid) {
 314                 printk(KERN_INFO
 315                        "node %d must be removed before remove section %ld\n",
 316                        nid, usemap_snr);
 317                 return;
 318         }
 319         /*
 320          * There is a circular dependency.
 321          * Some platforms allow un-removable section because they will just
 322          * gather other removable sections for dynamic partitioning.
 323          * Just notify un-removable section's number here.
 324          */
 325         printk(KERN_INFO "Section %ld and %ld (node %d)", usemap_snr,
 326                pgdat_snr, nid);
 327         printk(KERN_CONT
 328                " have a circular dependency on usemap and pgdat allocations\n");
 329 }
 330 #else
 331 static unsigned long * __init
 332 sparse_early_usemap_alloc_pgdat_section(struct pglist_data *pgdat)
 333 {
 334         return NULL;
 335 }
 336
 337 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 338 {
 339 }
 340 #endif /* CONFIG_MEMORY_HOTREMOVE */
 341
 342 static unsigned long *__init sparse_early_usemap_alloc(unsigned long pnum)
 343 {
 344         unsigned long *usemap;
 345         struct mem_section *ms = __nr_to_section(pnum);
 346         int nid = sparse_early_nid(ms);
 347
 348         usemap = sparse_early_usemap_alloc_pgdat_section(NODE_DATA(nid));
 349         if (usemap)
 350                 return usemap;
 351
 352         usemap = alloc_bootmem_node(NODE_DATA(nid), usemap_size());
 353         if (usemap) {
 354                 check_usemap_section_nr(nid, usemap);
 355                 return usemap;
 356         }
 357
 358         /* Stupid: suppress gcc warning for SPARSEMEM && !NUMA */
 359         nid = 0;
 360
 361         printk(KERN_WARNING "%s: allocation failed\n", __func__);
 362         return NULL;
 363 }
 364
 365 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 366 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
 367 {
 368         struct page *map;
 369
 370         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 371         if (map)
 372                 return map;
 373
 374         map = alloc_bootmem_pages_node(NODE_DATA(nid),
 375                        PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION));
 376         return map;
 377 }
 378 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 379
 380 struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 381 {
 382         struct page *map;
 383         struct mem_section *ms = __nr_to_section(pnum);
 384         int nid = sparse_early_nid(ms);
 385
 386         map = sparse_mem_map_populate(pnum, nid);
 387         if (map)
 388                 return map;
 389
 390         printk(KERN_ERR "%s: sparsemem memory map backing failed "
 391                         "some memory will not be available.\n", __func__);
 392         ms->section_mem_map = 0;
 393         return NULL;
 394 }
 395
 396 void __attribute__((weak)) __meminit vmemmap_populate_print_last(void)
 397 {
 398 }
 399 /*
 400  * Allocate the accumulated non-linear sections, allocate a mem_map
 401  * for each and record the physical to section mapping.
 402  */
 403 void __init sparse_init(void)
 404 {
 405         unsigned long pnum;
 406         struct page *map;
 407         unsigned long *usemap;
 408         unsigned long **usemap_map;
 409         int size;
 410
 411         /*
 412          * map is using big page (aka 2M in x86 64 bit)
 413          * usemap is less one page (aka 24 bytes)
 414          * so alloc 2M (with 2M align) and 24 bytes in turn will
 415          * make next 2M slip to one more 2M later.
 416          * then in big system, the memory will have a lot of holes...
 417          * here try to allocate 2M pages continously.
 418          *
 419          * powerpc need to call sparse_init_one_section right after each
 420          * sparse_early_mem_map_alloc, so allocate usemap_map at first.
 421          */
 422         size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
 423         usemap_map = alloc_bootmem(size);
 424         if (!usemap_map)
 425                 panic("can not allocate usemap_map\n");
 426
 427         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 428                 if (!present_section_nr(pnum))
 429                         continue;
 430                 usemap_map[pnum] = sparse_early_usemap_alloc(pnum);
 431         }
 432
 433         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 434                 if (!present_section_nr(pnum))
 435                         continue;
 436
 437                 usemap = usemap_map[pnum];
 438                 if (!usemap)
 439                         continue;
 440
 441                 map = sparse_early_mem_map_alloc(pnum);
 442                 if (!map)
 443                         continue;
 444
 445                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 446                                                                 usemap);
 447         }
 448
 449         vmemmap_populate_print_last();
 450
 451         free_bootmem(__pa(usemap_map), size);
 452 }
 453
 454 #ifdef CONFIG_MEMORY_HOTPLUG
 455 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 456 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 457                                                  unsigned long nr_pages)
 458 {
 459         /* This will make the necessary allocations eventually. */
 460         return sparse_mem_map_populate(pnum, nid);
 461 }
 462 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 463 {
 464         return; /* XXX: Not implemented yet */
 465 }
 466 static void free_map_bootmem(struct page *page, unsigned long nr_pages)
 467 {
 468 }
 469 #else
 470 static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
 471 {
 472         struct page *page, *ret;
 473         unsigned long memmap_size = sizeof(struct page) * nr_pages;
 474
 475         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 476         if (page)
 477                 goto got_map_page;
 478
 479         ret = vmalloc(memmap_size);
 480         if (ret)
 481                 goto got_map_ptr;
 482
 483         return NULL;
 484 got_map_page:
 485         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 486 got_map_ptr:
 487         memset(ret, 0, memmap_size);
 488
 489         return ret;
 490 }
 491
 492 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 493                                                   unsigned long nr_pages)
 494 {
 495         return __kmalloc_section_memmap(nr_pages);
 496 }
 497
 498 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 499 {
 500         if (is_vmalloc_addr(memmap))
 501                 vfree(memmap);
 502         else
 503                 free_pages((unsigned long)memmap,
 504                            get_order(sizeof(struct page) * nr_pages));
 505 }
 506
 507 static void free_map_bootmem(struct page *page, unsigned long nr_pages)
 508 {
 509         unsigned long maps_section_nr, removing_section_nr, i;
 510         int magic;
 511
 512         for (i = 0; i < nr_pages; i++, page++) {
 513                 magic = atomic_read(&page->_mapcount);
 514
 515                 BUG_ON(magic == NODE_INFO);
 516
 517                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 518                 removing_section_nr = page->private;
 519
 520                 /*
 521                  * When this function is called, the removing section is
 522                  * logical offlined state. This means all pages are isolated
 523                  * from page allocator. If removing section's memmap is placed
 524                  * on the same section, it must not be freed.
 525                  * If it is freed, page allocator may allocate it which will
 526                  * be removed physically soon.
 527                  */
 528                 if (maps_section_nr != removing_section_nr)
 529                         put_page_bootmem(page);
 530         }
 531 }
 532 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 533
 534 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
 535 {
 536         struct page *usemap_page;
 537         unsigned long nr_pages;
 538
 539         if (!usemap)
 540                 return;
 541
 542         usemap_page = virt_to_page(usemap);
 543         /*
 544          * Check to see if allocation came from hot-plug-add
 545          */
 546         if (PageSlab(usemap_page)) {
 547                 kfree(usemap);
 548                 if (memmap)
 549                         __kfree_section_memmap(memmap, PAGES_PER_SECTION);
 550                 return;
 551         }
 552
 553         /*
 554          * The usemap came from bootmem. This is packed with other usemaps
 555          * on the section which has pgdat at boot time. Just keep it as is now.
 556          */
 557
 558         if (memmap) {
 559                 struct page *memmap_page;
 560                 memmap_page = virt_to_page(memmap);
 561
 562                 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 563                         >> PAGE_SHIFT;
 564
 565                 free_map_bootmem(memmap_page, nr_pages);
 566         }
 567 }
 568
 569 /*
 570  * returns the number of sections whose mem_maps were properly
 571  * set.  If this is <=0, then that means that the passed-in
 572  * map was not consumed and must be freed.
 573  */
 574 int sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
 575                            int nr_pages)
 576 {
 577         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 578         struct pglist_data *pgdat = zone->zone_pgdat;
 579         struct mem_section *ms;
 580         struct page *memmap;
 581         unsigned long *usemap;
 582         unsigned long flags;
 583         int ret;
 584
 585         /*
 586          * no locking for this, because it does its own
 587          * plus, it does a kmalloc
 588          */
 589         ret = sparse_index_init(section_nr, pgdat->node_id);
 590         if (ret < 0 && ret != -EEXIST)
 591                 return ret;
 592         memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages);
 593         if (!memmap)
 594                 return -ENOMEM;
 595         usemap = __kmalloc_section_usemap();
 596         if (!usemap) {
 597                 __kfree_section_memmap(memmap, nr_pages);
 598                 return -ENOMEM;
 599         }
 600
 601         pgdat_resize_lock(pgdat, &flags);
 602
 603         ms = __pfn_to_section(start_pfn);
 604         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 605                 ret = -EEXIST;
 606                 goto out;
 607         }
 608
 609         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 610
 611         ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 612
 613 out:
 614         pgdat_resize_unlock(pgdat, &flags);
 615         if (ret <= 0) {
 616                 kfree(usemap);
 617                 __kfree_section_memmap(memmap, nr_pages);
 618         }
 619         return ret;
 620 }
 621
 622 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms)
 623 {
 624         struct page *memmap = NULL;
 625         unsigned long *usemap = NULL;
 626
 627         if (ms->section_mem_map) {
 628                 usemap = ms->pageblock_flags;
 629                 memmap = sparse_decode_mem_map(ms->section_mem_map,
 630                                                 __section_nr(ms));
 631                 ms->section_mem_map = 0;
 632                 ms->pageblock_flags = NULL;
 633         }
 634
 635         free_section_usemap(memmap, usemap);
 636 }
 637 #endif